About a year ago I participated in an art project for a hackathon where we wanted to have a swarm of "creatures" under computer control. The concept called for these creatures to participate in a simulated society, which would be guided by a crowd of humans online. They had to look like real animals, at least when viewed via a cheap webcam. We had a budget for materials, but it was fairly small and had to include many things other than robots.
For me it was an exciting challenge: hack together a capable robot swarm on a shoestring budget and on a hackathon's timeframe. I wanted to answer the question: What does it take to get a bunch of robots moving around on a surface under the control of a computer? To get started I went looking for prior art, and found the following representative projects:
- Kilobots (story) - Move on little stilt legs by vibration. Communicate with each other using reflected light. Various sources claim $14 per robot but commercially it costs $1127.28 for a pack of 10 robots ($112.73 per robot).
- AERobot - Can be purchased from Seeed Studio for $20 per robot. Uses vibration for motion, plugs directly into USB port.
- Jasmine - ~$120 per robot.
- mROBerTO - From University of Toronto.
- Droplet - From University of Colorado Boulder. Nice open design.
- R-One - From Rice University.
- Zooids - From Stanford University - design is open source.
- Actuated Workbench - Uses magnets in a surface to actuate a swarm of magnetic objects (nice quote: "...the actuated workbench works even when set on fire").
- Madgets - Also uses magnets under a surface to actuate objects above.
- Tangible Bots - Robots used as interface elements on tabletop display (paper).
- Thumbles - Swarm of robots with omni-wheels used for tabletop display interfaces.
- BitDrones - Drone cubes for 3D user interfaces. 3 types: 1) Pixel Drones (w/ OLED display) 2) Shape Drones (cube) 3) Display Drones (big display).
- "Image and Animation Display with Multiple Mobile Robots".
All of these projects are fascinating and accomplish different goals using a swarm of robots. However they are all similar in that they all take advantage of expensive custom hardware to achieve their research goals. In one case, the Kilobots project, a claimed goal is affordable hardware ($14/robot claimed in this Fast Company article), but the real cost is much higher ($112.73/robot).
I wanted to make a truly cheap swarm. So I looked to consumer electronics and searched for devices that had the basic capabilities I needed to serve as a starting point. That led me to the Hexbug Fire Ant toy.
Each one of these toys has tank-style driving remote-controlled by an infrared controller. They can move forward and backward and turn in either direction. Unfortunately the communications are not separated into channels, so any robot will react to the signals sent by any controller. All I wanted for my base was wireless control and the ability to move anywhere on a 2D surface, so the Hexbugs were enough to go on. I got 28 to serve as raw material for my swarm. Each cost <$20 at the time, though at the time of writing the price has risen to twice that.
Reverse Engineering the Communication Protocol
The first thing I did when I got the toys (after driving them all over my floor) was try to figure out how the infrared remote control works so I could control a robot from my computer. I broke open a controller and connected a probe to its IR LED. I also broke open a robot and soldered a wire to the output of its IR sensor. Then I drove the robot around with the controller while capturing with my logic analyzer and saw traces like the following:
There's a 38 KHz carrier and a simple short/long bit encoding scheme. Every message starts with an extra long pulse. And then one can choose to interpret short pulses as zeroes and the longer pulses as ones. The last 6 bits of the message vary between commands.
I pushed all the buttons, recorded the messages sent, and manually decoded them into this list of available commands:
- STOP = 0b000001
- FORWARD = 0b000111
- BACKWARD = 0b011001
- LEFT FORWARD = 0b000011
- LEFT BACKWARD = 0b010001
- RIGHT FORWARD = 0b000101
- RIGHT BACKWARD = 0b001001
Then I slapped together my own controller using a Raspberry Pi, an AVR microcontroller, and a large IR emitter. I called it the Horus Eye after the Egyptian god of the sky (among other things).
I included a camera because I wanted to eventually use it to track the robots for closed loop control. But to get started I wrote a Python program to control a single robot using a PS3 controller:
Hacking the Hardware
All of this initial work gave me a basis for experimentation, but it did not get me a swarm. If every robot responds to every command then there is no way to individually control them. To achieve that I would need to modify the toys. And to be able to modify them I needed to understand how they work. So I set about dissecting one.
When reverse engineering hardware my first step is often to take it apart and take the best pictures I can of the top and bottom of the PCB. I bring the photos into Photoshop and use the Warp Perspective tool to precisely line up the two sides of the PCB. Then I carefully trace the copper on each side to produce a map of the circuit that I can refer back to while I work out exactly what everything does.
There are very few parts in the control circuit for the toy. The main parts are the visible-light LED, two motor driver ICs, a glob top IC, a button pad, and an IR sensor. Everything is powered by a stack of coin cell batteries. Two DC motors drive the wheels of the robot via a gear train.
I did not see a way to convince the controller under the glob top to control the toy any differently, so I decided to start my modifications by removing it. If you are lucky these epoxy blobs can be cleanly removed with careful application of some flush cutters.
With the original controller gone I needed to add my own to get the toy to do anything again. When I was tracing the PCB I noted that the pinout of the motor drivers conveniently placed VCC and ground on the same pins as an ATtiny10 in its small outline transistor (SOT) package. The ATtiny10 in the SOT package has only 4 I/O pins, and one of those is the RESET line by default.
If I just wanted to drive the robot around then 4 pins is plenty because each motor driver uses 2. But I also want to receive communications from the outside world, so I probably need at least one I/O pin for that. To overcome this I decided to sacrifice some mobility. If I lost the ability to reverse on the right side then I could free up a pin for communications but still drive anywhere if given some extra time to spin all the way around. Because of the pin mapping the modification turned out to be fairly clean, requiring the microcontroller and just two wires.
It's a little hard to see but there is an ATtiny10 riding the the motor driver on one side. A short wire jumps over to the motor driver on the opposite side. And a longer wire jumps to the demodulated output of the IR sensor.
The parasite successfully infected the host. The toy is a zombie under its control.
I wrote some initial basic firmware that just turned the motor control pins on and off over and over to see that the microcontroller was working as expected. Then the real work began: I needed to implement a new communication protocol that would allow every robot in a swarm to be individually controlled.
The Zombie Language
All that was needed was a way for a given robot to receive the same control commands as the original toy, but know whether or not it should follow them based on its unique ID. Then the remote control system can individually control every robot in the swarm by giving them commands one by one.
One of the first questions is: how much data should be sent for each command? We need to address the robot we want to command and tell it what we want it to do, so there will be an address and command portion in every message. There are 4 commands (2 bits) but maybe I will want to add a few later, so I need at least 3 bits for them. I wanted to support a reasonable swarm. I had 28 robots to control. So I chose 5 bits for the address, to let me support swarms up to 32 robots in size. 3 bits + 5 bits = 1 byte, which is nice. That's the size of the registers on the AVR CPU in the ATtiny10, so that size will make it easier to write the code for handling communications.
I ended up writing the firmware in assembly. You can find it on GitHub here. But I'll go over the most significant parts of the code here. Let's look at initialization and the main loop first:
; Execution starts here reset: ; Set outputs for controlling motors sbi DDR_MOTOR, PIN_REV_L sbi DDR_MOTOR, PIN_FWD_L sbi DDR_MOTOR, PIN_FWD_R ; Enable TIMER0 with 64 divider (15625/s at 1MHz) ldi r16, 0x3 out TCCR0B, r16 ; Enable PCINT0 (on PB0) ldi r16, 1 << PCINT0 out PCMSK, r16 ; Enable pin change interrupts ldi r16, 1 << PCIE0 out PCICR, r16 ; Enable interrupts sei loop: rjmp loop
During initialization the code configures the I/O pins we need for output, starts a hardware timer to use as our sensor-sampling clock, and enables the pin change interrupt for the sensor I/O pin. Then the main loop just does nothing forever. Everything the robot will do will happen when requested by a command that is received, so the real action will happen in the IR sensor interrupt handler.
.macro measure_pulse in r16, TCNT0L in r17, TCNT0H time_pulse_start_wait_%: sbis PINPORT_RX, PIN_RX rjmp time_pulse_start_wait_% ; Calculate pulse length in r18, TCNT0L in r19, TCNT0H ; r18:r19 is pulse length in clock1 ticks sub r18, r16 sbc r19, r17 .endm ; Receive IR packet ir_interrupt: ; Exit interrupt if we are not here because of a falling edge sbic PINB, PIN_RX reti ; Prevent further interrupts while we are handling this one cli time_pulse: ; Measures length of first pulse ; and skip if it is not a (long) start pulse ; 1.193ms (0x1F timer diff) is a long pulse measure_pulse ; Exit interrupt if this is not a long pulse ldi r16, low(PULSE_LONG) ldi r17, high(PULSE_LONG) cp r18, r16 cpc r19, r17 brlt ir_interrupt_done ; It's a start pulse! ir_interrupt_start: ; Sample the rest of the bits in the incoming message clr r20 ; For the received bits ldi r21, PACKET_BIT_COUNT ir_interrupt_read_bit: sbic PINPORT_RX, PIN_RX rjmp ir_interrupt_read_bit measure_pulse ldi r16, low(PULSE_MED) ldi r17, high(PULSE_MED) cp r18, r16 cpc r19, r17 brlt ir_interrupt_read_bit_0 ir_interrupt_read_bit_1: sec rjmp ir_interrupt_read_finish ir_interrupt_read_bit_0: clc ir_interrupt_read_finish: ror r20 ; Shift received bit in dec r21 brne ir_interrupt_read_bit ; We are done reading the packet! Received value in r21 mov r21, r20 ; Does the ID match ours? andi r20, 0x1F ; Mask the ID (lower 5 bits) cpi r20, MY_ID brne ir_interrupt_not_me ; (Nope) ; The command is addressed to us! ; Let's decode and execute it mov r20, r21 andi r20, 0xE0 ; Mask the command (upper 3 bits) cpi r20, CMD_STOP << 5 brne ir_interrupt_check_fwd_l rjmp ir_interrupt_done ir_interrupt_check_fwd_l: cpi r20, CMD_FORWARD_LEFT << 5 brne ir_interrupt_check_fwd_r rjmp ir_interrupt_done ir_interrupt_check_fwd_r: cpi r20, CMD_FORWARD_RIGHT << 5 brne ir_interrupt_check_rev_l rjmp ir_interrupt_done ir_interrupt_check_rev_l: cpi r20, CMD_REVERSE_LEFT << 5 brne ir_interrupt_done rjmp ir_interrupt_done ir_interrupt_not_me: ir_interrupt_done: ; Re-enable interrupts and resume doing nothing sei reti
Short and sweet, only a couple dozen instructions. When a message starts there is a long start bit which triggers the IR interrupt. The handler then starts measuring the lengths of the incoming pulses (using the measure_pulse macro). Short pulses are interpreted as zeroes and long pulses are interpreted as ones. Once all 8 bits have been received the address is checked against that of the robot. If it matches then the command is decoded and executed, changing the states of the I/O pins to drive or stop the motor on each side of the robot.
Raising a Zombie Swarm
So now there is a way to tell individual robots what to do. The rest is easy, right? All we have to do is modify the rest of the toys and we have our robot swarm. Easy.
If you have ever done short-run electronics manufacturing you know how this goes. You pour effort into a beautiful design, make a working prototype, sit back and look at it running on your desk, and bask for a fleeting moment in the thought, "ah, my work is finally done." Then you blink, and free fall down out of the bright sunny space above the clouds into the cold black abyss of reality. You don't just need one, you need 10 or 100 or 1000. Your work has just begun.
So it was for me building my robot swarm. For each of the 28 robots I carefully flashed the microcontroller with the firmware containing the robot's unique ID, chopped off the top of the toy, popped the glob top off with clippers, applied the modification under a microscope with tweezers and soldering iron, tested it, and fixed any issues if necessary. It probably took a solid 8 hours to finish, but in the end I had the bodies for my swarm:
With the hardware ready I set about the next phase of the project, coordinating the swarm and putting them under precise control.
Controlling the Swarm
As a simple test I extended my PS3 controller code to let me select individual robots to control.
It was fun to drive them around this way, and for a moment I considered adding support for multiple controllers and racing robots with my friends. But time for the project was running out. A human controlling a single robot would not do. I needed a crowd of humans to control a swarm of robots.
Controlling this swarm is challenging because of a couple of factors. Communication is in one direction, from the control system to the robots, so it's not possible to ask the robots about their states. The robots don't have sensors besides the IR receiver so they do not have anything interesting to say anyway. So although the robot with a given ID can be told to move in any direction, without information about where it is and what direction it is facing it's not really possible to get it to a specific location. That's where the camera on the Eye of Horus comes in.
With a camera mounted above the space the swarm occupies it's possible to figure out where robots are and where they are facing by telling random robots to move and watching what parts of the image change. I used a technique called optical flow analysis implemented in Python using the computer vision library OpenCV to find regions of the video with movement and identify what direction the movement happened in. Once I had that information I was able to label robots and track their positions and orientations over time.
The next step would have been to write a motion planner to make it possible to specify a path for each individual robot to follow. But the project deadline loomed large and we had to cut things short to complete the art installation. I used what I had to link activity in the artificial society simulation (code here) to the activity of random robots. It turned out to be enough to produce relatively lifelike movement.
The Art Installation
To complete the illusion we were aiming for, of a group of animals controlled by the aggregate activity of humans, we covered each of the robots in hair:
Viewed from above via webcam they actually looked fairly convincing. The movements were a little more like those of insects than furry mammals, but it was close enough for us.
For the final setup we put the robots in an aquarium with plants and rocks to mimic a natural environment.
So ends the saga of the swarm. But you can read about the overall art project that this was a part of in this article on the MIT CAST website.
If you found this interesting and want to see other things that I'm working on I often post pictures and videos on my Instagram. And feel free to send me an email if you have feedback, interesting ideas, or want to collaborate.